Scholarlink

This paper presents a global path planning method based on the Deep Double Q Network (DDQN) to improve the autonomous navigation capability of amphibious unmanned aerial vehicles (AUSVs) in complex environments. First, a simulation environment is constructed using a generalized super-ellipsoid model. Second, a reinforcement learning framework is designed, including state space, action space, and reward functions tailored for time sensitive and energy sensitive tasks. An action masking mechanism is also applied to filter out infeasible actiony. Finally, a third-order Bézier curve is applied to smooth the generated path, ensuring better compliance with the motion characteristics of AUSVs. Simulation results demonstrate that the proposed method can generate suitable paths based on different task requirements and outperforms the rapidly-exploring random tree (RRT) algorithm in terms of planning performance, validating its effectiveness and superiority.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Reinforcement Learning-Based Multi-objective Global Path Planning for Amphibious Unmanned Surface Vehicles

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dong Wu,
  • Ronghao Wang,
  • Shuo Li,
  • Zhengrong Xiang

摘要

This paper presents a global path planning method based on the Deep Double Q Network (DDQN) to improve the autonomous navigation capability of amphibious unmanned aerial vehicles (AUSVs) in complex environments. First, a simulation environment is constructed using a generalized super-ellipsoid model. Second, a reinforcement learning framework is designed, including state space, action space, and reward functions tailored for time sensitive and energy sensitive tasks. An action masking mechanism is also applied to filter out infeasible actiony. Finally, a third-order Bézier curve is applied to smooth the generated path, ensuring better compliance with the motion characteristics of AUSVs. Simulation results demonstrate that the proposed method can generate suitable paths based on different task requirements and outperforms the rapidly-exploring random tree (RRT) algorithm in terms of planning performance, validating its effectiveness and superiority.